AI NEWS REPORT

Research post and GitHub repo, surfaced on Anthropic's site this week. Two Anthropic staff with biology backgrounds and no kernel-engineering experience supervised Claude for about four weeks. Claude optimized more than 30 open-source models for structure prediction, protein design, genomics and sequence modeling. Average speedup about 4x with small precision loss, about 1.6x with identical outputs. New FlashPairformer kernels run triangle attention 2.7 to 2.9x faster and triangle multiplication 1.7 to 3.2x faster than the field standard. A 'big mode' lets a single NVIDIA node predict structures over 10,000 tokens, which was not possible before, and a 70,000-token bacterial ribosome ran on one B300 node. Anthropic says a binder design campaign that used to take thousands of H100 hours cost about $150 in GPU and tokens. All the code is in the repo. Why it matters: it is a concrete, checkable example of an agent doing weeks of specialist engineering, and the kernels are free to use. What to watch: the speedups are Anthropic's measurements. Independent labs have not replicated them yet.

Engineering post, September 21, 282 points and 340 comments on Hacker News. The thesis: 'Agents have made it exponentially faster to ship code, but validating those changes hasn't quite kept up.' Linear's test suite nearly quadrupled since January. The fixes, with numbers: moving off GitHub Actions to third-party runners made jobs 34 percent faster; the tsgo TypeScript compiler cut type-checking 73 percent; change detection went from 26 to 8 seconds; pre-installing dependencies in CI images took setup from 44 seconds to 16 to 18; going from 4 to 8 shards and sharing module state with strict isolation saved 17 percent, the single biggest win. One optimization alone saves about 87,000 runner-minutes a month, 11.8 percent of their CI. PR wait time went from over 6 minutes to just over 5. Why it matters: if your team turned on coding agents this year, your CI bill and your merge queue are about to look like this. Steal the list.

The New Stack, September 19, by Arjun Iyer, CEO of Signadot. Disclosure first: Signadot sponsored the post and it ends in a pitch for Signadot's Kubernetes product. Read it anyway, because the first two thirds are a clear argument. It starts from Lauren Tan, an engineer on SpaceXAI's Grok team, formerly at Cursor and Meta, who published her personal agent workflow, pstack, and says it lets her ship about 2,000 pull requests a month to production. Her key piece is a verification skill: a CLI the agent uses to start the app, drive it, inspect state and read structured JSON back, so it keeps fixing until the checks pass. Her line: 'An agent that can check its own output keeps working until the task is done. An agent that can't hand you a diff and wait makes you the slowest component in the loop.' The math: at 2,000 PRs a month a human gets about five minutes per PR, so review cannot be the verification layer. Iyer's point is that Tan's trick works because her app fits in one process. In a system of sixty services, a CLI can start the changed service but not the system. He walks through why the three usual runtimes fail at agent scale: mocks drift so the agent 'closes its loop against fiction,' a full stack per change costs too much and takes minutes, and shared staging lets hundreds of agents overwrite each other. Then five requirements: real dependencies, isolation between concurrent changes, cost that scales with the change not the system, environments in seconds, and access through the agent's own CLI or MCP. The proposed answer, environments as a view of one shared stable stack with only the changed service swapped in, is the Signadot product. Why it matters: it is the sharpest statement this month of the rule that evidence has to come before a human spends time on agent output. What to watch: the 2,000 figure is Tan's own claim about her own workflow, and the closing section is an ad.

IndyDevDan, September 21. 'Self-Compact Pi Agent: ZERO HYPE Agentic Coding Devlog.' A working engineer's take on context compaction in a real agent, the same problem Tim Dettmers's CliffCompaction claims to solve. Pairs with today's Frontier.

IndyDevDan · 21.8K views · 375 likes

AI Engineer, September 21. Paul Bakaus, 'The Dark Arts of Skill Engineering.' How to write the skill files that agents load, and where they go wrong. For anyone maintaining a CLAUDE.md or AGENTS.md.

AI Engineer · 15.1K views · 126 likes

Featured by Latent Space.

AI Engineer

Featured by Latent Space.

Salesforce

Featured by Latent Space.

Matija Grcic

Posted September 21. Willison explains TypeSafe AI's Jev, the 'decision model' behind this week's Kev clones and the Berman and Medin videos. You send a state object and questions; you get back probabilities: yes or no, a pick from a list, or a score on a scale, many questions per request, in parallel, with no text at all. Input costs $0.042 per million tokens and output is free. He likes it for spam filtering, labeling, ranking and reranking search results. His warning is the point of the post: a normal model at least shows you its words, but 'the only thing you're going to get back is a floating point number,' so hidden bias is invisible, and he says do not point it at hiring. His rule: 'evals and structured experiments are even more important than they are for regular LLM projects.' Why read it: the clearest explanation of the new model type in about 800 words, from someone who has already tried it.

Published September 17, four days before the release, and the best background read on today's siren. Xiaomi opened a real-time dashboard on September 15 showing cost, tokens and benchmark scores while the RL run was still going: about $432,000 a day for Pro, 32.5 billion tokens processed, DeepSWE climbing from 19 to 66 percent. Luo Fuli, who leads the team: 'Nearly half a year of silence. We spent it studying one problem: how far RL can scale.' The same dashboard carried a line labeled 'Claude Distill Requests: hidden,' which readers spotted and which Xiaomi has not explained. Why read it: it is the only account of the training run written before anyone knew how it would end, and it names the question the launch post does not.